Current artificial systems suffer from catastrophic forgetting during continual learning, a limitation absent in biological systems. Biological mechanisms leverage the dual representation of specific and generalized memories within corticohippocampal circuits to facilitate lifelong learning. Inspired by this, we develop a corticohippocampal circuits-based hybrid neural network (CH-HNN) that emulates these dual representations, significantly mitigating catastrophic forgetting in both task-incremental and class-incremental learning scenarios. Our CH-HNNs incorporate artificial neural networks and spiking neural networks, leveraging prior knowledge to facilitate new concept learning through episode inference, and offering insights into the neural functions of both feedforward and feedback loops within corticohippocampal circuits. Crucially, CH-HNN operates as a task-agnostic system without increasing memory demands, demonstrating adaptability and robustness in real-world applications. Coupled with the low power consumption inherent to SNNs, our model represents the potential for energy-efficient, continual learning in dynamic environments.

Spiking neural networks (SNNs) excel in spatio-temporal information processing, yet computational intensity limits their adoption. Although prior research have focused on spatial pruning or neuronal dynamics optimization for energy-efficiency, pruning based on neuronal dynamics parameters remains unexplored. This paper introduces a two-stage evolutionary compression technique that utilizes rich dynamics patterns to guide the pruning process, termed Spiking Neuronal Dynamics Pruning (SNDP). SNDP comprises Neuronal Dynamics Searcher (NDS) and Dynamics Pattern Transducer (DPT). NDS uses an evolutionary strategy to identify neuron clusters with optimal dynamics for achieving high performance. DPT then prunes neurons by retaining one from each cluster, preserving key parameters and synaptic weights identified by NDS. Experiments show that SNDP reduces inference energy by up to 97.5% in various SNNs, with minimal performance loss across diverse neuromorphic vision and speech datasets. Furthermore, pruned networks show enhanced noise robustness compared to same-sized traditional trained SNNs. SNDP shows promise in compressing SNNs by harnessing the power of neural dynamics.

In recent years, spiking neural networks (SNNs) have attracted substantial interest due to their potential to replicate the energy-efficient and event-driven processing of biological neurons. Despite this, the application of SNNs in graph representation learning, particularly for non-Euclidean data, remains underexplored, and the influence of spiking dynamics on graph learning is not yet fully understood. This work seeks to address these gaps by examining the unique properties and benefits of spiking dynamics in enhancing graph representation learning. We propose a spike-based graph neural network model that incorporates spiking dynamics, enhanced by a novel spatial-temporal feature normalization (STFN) technique, to improve training efficiency and model stability. Our detailed analysis explores the impact of rate coding and temporal coding on SNN performance, offering new insights into their advantages for deep graph networks and addressing challenges such as the oversmoothing problem. Experimental results demonstrate that our SNN models can achieve competitive performance with state-of-the-art graph neural networks (GNNs) while considerably reducing computational costs, highlighting the potential of SNNs for efficient neuromorphic computing applications in complex graph-based scenarios.

Synaptic plasticity plays a critical role in the expression power of brain neural networks. Among diverse plasticity rules, synaptic scaling presents indispensable effects on homeostasis maintenance and synaptic strength regulation. In the current modeling of brain-inspired spiking neural networks (SNN), backpropagation through time is widely adopted because it can achieve high performance using a small number of time steps. Nevertheless, the synaptic scaling mechanism has not yet been well touched. In this work, we propose an experience-dependent adaptive synaptic scaling mechanism (AS-SNN) for spiking neural networks. The learning process has two stages: First, in the forward path, adaptive short-term potentiation or depression is triggered for each synapse according to afferent stimuli intensity accumulated by presynaptic historical neural activities. Second, in the backward path, long-term consolidation is executed through gradient signals regulated by the corresponding scaling factor. This mechanism shapes the pattern selectivity of synapses and the information transfer they mediate. We theoretically prove that the proposed adaptive synaptic scaling function follows a contraction map and finally converges to an expected fixed point, in accordance with state-of-the-art results in three tasks on perturbation resistance, continual learning, and graph learning. Specifically, for the perturbation resistance and continual learning tasks, our approach improves the accuracy on the N-MNIST benchmark over the baseline by 44% and 25%, respectively. An expected firing rate callback and sparse coding can be observed in graph learning. Extensive experiments on ablation study and cost evaluation evidence the effectiveness and efficiency of our nonparametric adaptive scaling method, which demonstrates the great potential of SNN in continual learning and robust learning.

Brain-inspired computing, drawing inspiration from the fundamental structure and information-processing mechanisms of the human brain, has gained significant momentum in recent years. It has emerged as a research paradigm centered on brain-computer dual-driven and multi-network integration. One noteworthy instance of this paradigm is the hybrid neural network (HNN), which integrates computer-science-oriented artificial neural networks (ANNs) with neuroscience-oriented spiking neural networks (SNNs). HNNs exhibit distinct advantages in various intelligent tasks, including perception, cognition and learning. This paper presents a comprehensive review of HNNs with an emphasis on their origin, concepts, biological perspective, construction framework and supporting systems. Furthermore, insights and suggestions for potential research directions are provided aiming to propel the advancement of the HNN paradigm.

Recent advances in artificial intelligence have enhanced the abilities of mobile robots in dealing with complex and dynamic scenarios. However, to enable computationally intensive algorithms to be executed locally in multitask robots with low latency and high efficiency, innovations in computing hardware are required. Here, we report TianjicX, a neuromorphic computing hardware that can support true concurrent execution of multiple cross-computing-paradigm neural network (NN) models with various coordination manners for robotics. With spatiotemporal elasticity, TianjicX can support adaptive allocation of computing resources and scheduling of execution time for each task. Key to this approach is a high-level model, “Rivulet,” which bridges the gap between robotic-level requirements and hardware implementations. It abstracts the execution of NN tasks through distribution of static data and streaming of dynamic data to form the basic activity context, adopts time and space slices to achieve elastic resource allocation for each activity, and performs configurable hybrid synchronous-asynchronous grouping. Thereby, Rivulet is capable of supporting independent and interactive execution. Building on Rivulet with hardware design for realizing spatiotemporal elasticity, a 28-nanometer TianjicX neuromorphic chip with event-driven, high parallelism, low latency, and low power was developed. Using a single TianjicX chip and a specially developed compiler stack, we built a multi-intelligent-tasking mobile robot, Tianjicat, to perform a cat-and-mouse game. Multiple tasks, including sound recognition and tracking, object recognition, obstacle avoidance, and decision-making, can be concurrently executed. Compared with NVIDIA Jetson TX2, latency is substantially reduced by 79.09 times, and dynamic power is reduced by 50.66%.

There is a growing trend to design hybrid neural networks (HNNs) by combining spiking neural networks and artificial neural networks to leverage the strengths of both. Here, we propose a framework for general design and computation of HNNs by introducing hybrid units (HUs) as a linkage interface. The framework not only integrates key features of these computing paradigms but also decouples them to improve flexibility and efficiency. HUs are designable and learnable to promote transmission and modulation of hybrid information flows in HNNs. Through three cases, we demonstrate that the framework can facilitate hybrid model design. The hybrid sensing network implements multi-pathway sensing, achieving high tracking accuracy and energy efficiency. The hybrid modulation network implements hierarchical information abstraction, enabling meta-continual learning of multiple tasks. The hybrid reasoning network performs multimodal reasoning in an interpretable, robust and parallel manner. This study advances cross-paradigm modeling for a broad range of intelligent tasks.

As the representatives of brain-inspired models at the neuronal level, spiking neural networks (SNNs) have shown great promise in processing spatiotemporal information with intrinsic temporal dynamics. SNNs are expected to further improve their robustness and computing efficiency by introducing top-down attention at the architectural level, which is crucial for the human brain to support advanced intelligence. However, this attempt encounters difficulties in optimizing the attention in SNNs largely due to the lack of annotations. Here, we develop a hybrid network model with a top-down attention mechanism (HTDA) by incorporating an artificial neural network (ANN) to generate attention maps based on the features extracted by a feedforward SNN. The attention map is then used to modulate the encoding layer of the SNN so that it focuses on the most informative sensory input. To facilitate direct learning of attention maps and avoid labor-intensive annotations, we propose a general principle and a corresponding weakly-supervised objective, which promotes the HTDA model to utilize an integral and small subset of the input to give accurate predictions. On this basis, the ANN and the SNN can be jointly optimized by surrogate gradient descent in an end-to-end manner. We comprehensively evaluated the HTDA model on object recognition tasks, which demonstrates strong robustness to adversarial noise, high computing efficiency, and good interpretability. On the widely-adopted CIFAR-10, CIFAR-100, and MNIST benchmarks, the HTDA model reduces firing rates by up to 50% and improves adversarial robustness by up to 10% with comparable or better accuracy compared with the state-of-the-art SNNs. The HTDA model is also verified on dynamic neuromorphic datasets and achieves consistent improvements. This study provides a new way to boost the performance of SNNs by employing a hybrid top-down attention mechanism.

Deep neural networks have achieved remarkable performance in various applications but are extremely vulnerable to adversarial perturbation. The most representative and promising methods that can enhance model robustness, such as adversarial training and its variants, substantially degrade model accuracy on benign samples, limiting practical utility. Although incorporating extra training data can alleviate the trade-off to a certain extent, it remains unsolved to achieve both robustness and accuracy under limited training data. Here, we demonstrate the feasibility of overcoming the trade-off, by developing an adversarial feature stacking (AFS) model, which combines multiple independent feature extractors with varied levels of robustness and accuracy. Theoretical analysis is further conducted, and general principles for the selection of basic feature extractors are provided. We evaluate the AFS model on CIFAR-10 and CIFAR-100 datasets with strong adaptive attack methods, significantly advancing the state-of-the-art in terms of the trade-off. The AFS model achieves a benign accuracy improvement of ∼ 6 ∼ 10

Plastic changes with intrinsic dynamics in synaptic efficacy underlie the cellular level of expression of brain functions regarding multimodal information processing. Among diverse plasticity mechanisms, synaptic scaling exerts indispensable effects on the homeostatic state maintenance and synaptic strength regulation in biological neural networks. Despite recent tremendous progress in developing spiking neural networks (SNNs) for multiple complex scenarios, most of the work remains in the pure backpropagation-based framework where the synaptic scaling mechanism is rarely effectively incorporated. In this work, we present a biologically inspired neuronal model with an activity-dependent adaptive synaptic scaling mechanism that endows each synapse with both short-term enhancement and depression properties. The learning process is completed in two phases. Firstly, in the forward conduction circuits, adaptive short-term enhancement or depression response is triggered in the light of afferent stimuli intensity; Then long-term consolidation is executed by back-propagated error signals. These processes dramatically shape the pattern selectivity of synapses and the diverse information transfer they mediate. Experiments reveal remarkable advantages in three tasks regarding bimodal learning. Specifically, On the continual learning and perturbation-resistant task for Dynamic Vision Sensor (DVS) modal information, our method improves the mean accuracy on the benchmark of N-MNIST dataset than the baseline by and , respectively. On sequence learning task for Active Pixel Sensor (APS) model information, our method improve the generalization capability and training stability by a large margin. These results demonstrate favourable effectiveness of such non-parametric adaptive strategy on bimodal information inference for APS and DVS, facilitating intelligence understanding and bio-inspired modelling.

Memristive devices based on two-dimensional (2D) semiconducting materials have emerged as highly promising neuromorphic devices due to their intrinsic atomic body and unique properties. However, the migration and redistribution of anions induces built-in electric field at 2D materials/electrode interface. It inevitably leads to nonlinearity and saturation of conductance change, which are the key challenges of 2D materials based synaptic devices to achieve high accuracy neuromorphic applications. In this work, we report a vertical heterostructure formed by monolayer CVD-grown MoS2 and WO3 films, in which the WO 3 films serve as anions reservoir to steadily absorb and release sulfur anions, thus successfully overcoming the hurdles of nonlinearity and limited conductance states. We experimentally demonstrate a nearly linear change in conductance (similar to 1.1) and as high as 130 (similar to 2(7)) weight states, which is a record among 2D materials-based synapses. Simulations prove that artificial neural network with MoS2/WO3 heterostructure synapses achieves a significantly improved learning accuracy of 93.2% in MNIST handwritten digits, demonstrating the dual benefits of linearity and multilevels caused by the anion reservoir. In addition, the essential synaptic behaviors, such as potentiation/depression, paired pulse facilitation, spike-rate-dependent plasticity as well as transformation from short-term plasticity to long-term plasticity are implemented in the heterostructure device. The introduction of anion reservoir opens an effective approach to overcome the limitations of 2D materials and enhance the performance of neuromorphic devices for high-precision neuromorphic computing.

Biological spiking neurons with intrinsic dynamics underlie the powerful representation and learning capabilities of the brain for processing multimodal information in complex environments. Despite recent tremendous progress in spiking neural networks (SNNs) for handling Euclidean-space tasks, it still remains challenging to exploit SNNs in processing non-Euclidean-space data represented by graph data, mainly due to the lack of effective modeling framework and useful training techniques. Here we present a general spike-based modeling framework that enables the direct training of SNNs for graph learning. Through spatial-temporal unfolding for spiking data flows of node features, we incorporate graph convolution filters into spiking dynamics and formalize a synergistic learning paradigm. Considering the unique features of spike representation and spiking dynamics, we propose a spatial-temporal feature normalization (STFN) technique suitable for SNN to accelerate convergence. We instantiate our methods into two spiking graph models, including graph convolution SNNs and graph attention SNNs, and validate their performance on three node-classification benchmarks, including Cora, Citeseer, and Pubmed. Our model can achieve comparable performance with the state-of-the-art graph neural network (GNN) models with much lower computation costs, demonstrating great benefits for the execution on neuromorphic hardware and prompting neuromorphic applications in graphical scenarios.

Deep neural networks have achieved remarkable performance in various applications but are extremely vulnerable to adversarial perturbation. The most representative and promising methods that can enhance model robustness, such as adversarial training and its variants, substantially degrade model accuracy on benign samples, limiting practical utility. Although incorporating extra training data can alleviate the trade-off to a certain extent, it remains unsolved to achieve both robustness and accuracy under limited training data. Here, we demonstrate the feasibility of overcoming the trade-off, by developing an adversarial feature stacking (AFS) model, which combines multiple independent feature extractors with varied levels of robustness and accuracy. Theoretical analysis is further conducted, and general principles for the selection of basic feature extractors are provided. We evaluate the AFS model on CIFAR-10 and CIFAR-100 datasets with strong adaptive attack methods, significantly advancing the state-of-the-art in terms of the trade-off. The AFS model achieves a benign accuracy improvement of ~6% on CIFAR-10 and ~10% on CIFAR-100 with comparable or even stronger robustness than the state-of-the-art adversarial training methods.

Generative adversarial networks have achieved remarkable performance on various tasks but suffer from sensitivity to hyper-parameters, training instability, and mode collapse. We find that this is partly due to gradient given by non-robust discriminator containing non-informative adversarial noise, which can hinder generator from catching the pattern of real samples. Inspired by defense against adversarial samples, we introduce adversarial training of discriminator on real samples that does not exist in classic GANs framework to make adversarial training symmetric, which can balance min-max game and make discriminator more robust. Robust discriminator can give more informative gradient with less adversarial noise, which can stabilize training and accelerate convergence. We validate the proposed method on image generation tasks with varied network architectures quantitatively. Experiments show that training stability, perceptual quality, and diversity of generated samples are consistently improved with small additional training computation cost.

Neuromorphic computing has attracted increasing attention in medical applications due to its ability to improve diagnosis accuracy and human healthcare monitoring. However, the current remote operation mode has a time delay between in vivo data acquisition and in vitro clinical decision-making. Thus, it is of great importance to build a biodegradable neuromorphic network that can operate in a local physiological environment. A biodegradable synapse is a crucial component of such neuromorphic networks. However, the materials employed currently to develop a biodegradable synapse are incompatible with the foundry process, making it challenging to achieve a high density and large-scale neuromorphic network. Here, we report a biodegradable artificial synapse based on a W/Cu/WO3/SiO2/W structure, which is constructed from materials widely used in advanced semiconductor foundries. The device exhibits resistive switching, and the dominated mechanisms are attributed to Ohm's law and trap-filled space charge limited conduction. By manipulating pulse amplitudes, widths, and intervals, the device conductance can be finely regulated to achieve various synaptic functions, such as long-term potentiation, long term depression, paired-pulse facilitation, and spike-rate-dependent plasticity. Moreover, the learning-forgetting-relearning process, which is an essential and complex synaptic behavior, is emulated in a single device. Pattern learning of a slash symbol is also accomplished by building a 4 × 4 synaptic array. In addition, the systematic solubility testing proves its full biodegradability in biofluids. This work opens a potential pathway toward the integration of large-scale neuromorphic network for bioelectronics.

Generative adversarial networks have achieved remarkable performance on various tasks but suffer from training instability. Despite many training strategies proposed to improve training stability, this issue remains as a challenge. In this paper, we investigate the training instability from the perspective of adversarial samples and reveal that adversarial training on fake samples is implemented in vanilla GANs, but adversarial training on real samples has long been overlooked. Consequently, the discriminator is extremely vulnerable to adversarial perturbation and the gradient given by the discriminator contains non-informative adversarial noises, which hinders the generator from catching the pattern of real samples. Here, we develop adversarial symmetric GANs (AS-GANs) that incorporate adversarial training of the discriminator on real samples into vanilla GANs, making adversarial training symmetrical. The discriminator is therefore more robust and provides more informative gradient with less adversarial noise, thereby stabilizing training and accelerating convergence. The effectiveness of the AS-GANs is verified on image generation on CIFAR-10, CIFAR-100, CelebA, and LSUN with varied network architectures. Not only the training is more stabilized, but the FID scores of generated samples are consistently improved by a large margin compared to the baseline. Theoretical analysis is also conducted to explain why AS-GAN can improve training. The bridging of adversarial samples and adversarial networks provides a new approach to further develop adversarial networks.